The ability to query a database privately is nowadays ubiquitous via an encrypted channel. With the advent of homomorphic encryption, there is a want to expand the notion of privacy in this context to querying privately on the database with the database learning as little to no information of the query data or its result. The ability to compute the intersection from at least two parties’ sets that are kept private only to themselves is known as private set intersection (PSI) and should be considered a fundamental operation in several homomorphic computation scenarios to do useful work; not least for the ability to implement queries on a database. We outline in this paper a novel highly configurable PSI structure to be used in private querying providing the possibility that even the exact query itself can be protected from the database if required. As well as complex database lookups, there is also a more complex partial matching. The outline of the system design is discussed and we report preliminary results on some of the fundamental operations. We demonstrate that this technology is emerging as a viable given response to lookup queries and partially matching on an encrypted database with over a million entries in approximately 9 minutes.

Edge devices are becoming increasingly pervasive in everyday life. These devices have become more computationally capable allowing for more complex AI reasoning at the edge. Subsequently, there is a need for protecting data to comply with privacy laws and confidentiality regulations. In this paper, we demonstrate the applicability of homomorphic encryption for protecting the data of AI-enabled cameras at the edge by implementing our solution on a commercial edge device. Our solution comprises a local homomorphic key-value database on an edge device populated by an AI camera; permitting the service of homomorphic search to be performed directly on the edge device. We characterize our implementation demonstrating linear behavior with respect to the database size that the edge device can support. Good enough performance is known to be difficult to achieve when employing homomorphic encryption. Our results are encouraging as we achieved solutions considered to be homomorphically fast, for example, linear performance of 1.28 seconds per database entry at over 256 bits of security. This amounts to a query being processed on a database of 200 entries in ~ 5 minutes.

Machine learning (ML) is today commonly employed in the Financial Services Sector (FSS) to create various models to predict a variety of conditions ranging from financial transactions fraud to outcomes of investments and also targeted marketing campaigns. The common ML technique used for the modeling is supervised learning using regression algorithms and usually involves large amounts of data that needs to be shared and prepared before the actual learning phase. Compliance with privacy laws and confidentiality regulations requires that most, if not all, of the data must be kept in a secure environment, usually in-house, and not outsourced to cloud or multi-tenant shared environments. This paper presents the results of a research collaboration between IBM Research and Banco Bradesco SA to investigate approaches to homomorphically secure a typical ML pipeline commonly employed in the FSS industry. We investigated and de-constructed a typical ML pipeline used by Banco Bradesco and applied Homomorphic Encryption (HE) to two of the important ML tasks, namely the variable selection phase of the model generation task and the prediction task. Variable selection, which usually precedes the training phase, is very important when working with data sets for which no prior knowledge of the covariate set exists. Our work provides a way to define an initial covariate set for the training phase while preserving the privacy and confidentiality of the input data sets. Quality metrics, using real financial data, comprising quantitative, qualitative and categorical features, demonstrated that our HE based pipeline can yield results comparable to state of the art variable selection techniques and the performance results demonstrated that HE technology has reached the inflection point where it can be useful in batch processing in a financial business setting.

We describe our recent experience, building a system that uses fully-homomorphic encryption (FHE) to approximate the coefficients of a logistic-regression model, built from genomic data. The aim of this project was to examine the feasibility of a solution that operates "deep within the bootstrapping regime,'' solving a problem that appears too hard to be addressed just with somewhat-homomorphic encryption. As part of this project, we implemented optimized versions of many bread and butter FHE tools. These tools include binary arithmetic, comparisons, partial sorting, and low-precision approximation of arbitrary functions (used for reciprocals, logarithms, etc.). Our solution can handle thousands of records and hundreds of fields, and it takes a few hours to run. To achieve this performance we had to be extremely frugal with expensive bootstrapping and data-movement operations. We believe that our experience in this project could serve as a guide for what is or is not currently feasible to do with fully-homomorphic encryption.